- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes Dirac Convolutional Networks (DCNs), a family of simplicial neural networks built from polynomial spectral filters of the Dirac operator, which recovers existing convolutional SNNs such as SCCNN and SCNN as special cases. The authors instantiate DCNs with the root-to-leaf normalized Dirac operator of Viganò et al. (2026) and Chebyshev polynomials of order two (R2L-DCNs). They show that a continuous linear version of the model with skew-symmetric weights preserves the norm and the Dirichlet energy, and that the forward-Euler discretization converges to the highest eigenvalue of the Hodge Laplacian, with rates that depend on the spectrum. The method is evaluated on trajectory prediction (ocean drifters), mesh classification (SHREC'11), node and graph classification, and a depth experiment on a rewired version of minesweeper.
Critical strengths
The paper is easy to follow, and the theoretical results and their proofs are clearly presented. Using a normalization that bounds the spectrum of the Dirac operator and all Hodge Laplacians in a common interval is a clean idea, and it makes Chebyshev filters natural in this setting. R2L-DCN obtains competitive results with the baselines, and Figure 4 shows that it keeps its performance when stacking many layers, whereas SCCNN degrades.
Critical weaknesses
I recommend rejection in its current form. The idea is clean, but the contribution is limited with respect to existing simplicial convolutional networks, and the main claims (transferability and stability) are not supported by the experiments or the theory.
- Limited novelty. Even powers of the Dirac operator are powers of the Hodge Laplacians, and odd powers are (co)boundary maps. With order two, R2L-DCN reduces to identity, up/down Hodge Laplacian, and incidence terms, i.e., an SCCNN layer with normalized operators. The main difference with SCCNN is therefore the normalization, which comes from Viganò et al. (2026). Moreover, Dirac-based polynomial filters for simplicial neural networks were already proposed in [R1], which isn't cited or discussed. The ablations also don't show a clear benefit of the inter-rank coupling: R2L-DCN(even), which removes it, is better than R2L-DCN on both irregular meshes in Table 1 (48.0 vs. 46.8 and 45.4 vs. 41.1).
- The main claims are not backed by experiments or theory.
- The abstract and the contributions claim that normalization allows filters to transfer across simplicial complexes, but there is no transferability experiment, and the conclusion lists "testing transfer to unseen complexes" as future work. Normalization guarantees that the spectra of different complexes lie in the same interval, but this isn't enough to claim that learned filters transfer: the informative eigenspaces can still be at different frequencies. The book-complex example in Figure 2 is a hand-picked illustration, not evidence.
- The theory doesn't directly apply to the proposed model. The analysis considers linear dynamics with only the first-order term and constrained (symmetric or skew-symmetric) weights, while R2L-DCN uses nonlinearities, unconstrained weights, and second-order Chebyshev terms. Moreover, Proposition 4.1 holds for any symmetric propagation operator with skew-symmetric weights, so it's unclear what is specific to the Dirac formulation.
- For the forward-Euler discretization with skew-symmetric weights, Eqn. (38) shows that every mode is amplified by a factor larger than one, so the norm grows with depth. Normalization only slows down this growth, which is weaker than the stability claimed in the abstract.
- There is no comparison between normalized and unnormalized DCNs with the same architecture, so the effect of R2L normalization on accuracy and depth robustness isn't isolated.
Minor: In Eqns. (16) and (40), the skew-symmetric rate uses , where is defined as an eigenvalue, which is purely imaginary in this case. It should be the squared modulus, as in Eqn. (38).
References
[R1] "Generalized simplicial attention neural networks", IEEE Transactions on Signal and Information Processing over Networks, 2024.
Questions for authors
Clarifications:
- Is R2L-DCN of order two equivalent to an SCCNN with R2L-normalized operators? If not, what can it represent that SCCNN can't? How is DCN positioned with respect to [R1]?
- To what extent do the theoretical results explain the behavior of R2L-DCN, given that the analysis focuses on linear dynamics with constrained weights? What is specific to the Dirac formulation in Proposition 4.1?
- In what precise sense is the Euler discretization stable, given that the norm grows with depth for skew-symmetric weights?
Suggestions for the revision:
- Provide experiments showing the transferability of R2L-DCNs across simplicial complexes, e.g., training on some complexes and testing on unseen ones with different sizes or degree distributions.
- Compare normalized and unnormalized DCNs with the same architecture to isolate the effect of R2L normalization on accuracy and depth robustness.
8 comments
Official Comment by Authors - Part 2:
Filter Transferability
The abstract and the contributions claim that normalization allows filters to transfer across simplicial complexes, but there is no transferability experiment,
We appreciate this comment, and now conduct experiments to address reviewer's concern. In particular, we design an edge-classification task, where the goal is to detect which edges are corrupted via in the existence of a gradient signal coming from the vertices, where the filters are trained on a small (60 vertices) simplicial complex and transferred to larger complexes. We added Section 5.3 and Appendix D.2 for further details and discussion of the problem. We also discuss the random-walk variant of R2L upon the suggestion of another reviewer.
Table A: Balanced accuracy (%) in corrupted-edge detection task for filters of order K.
| Operator | K | n=60 | n=240 | Irreg. n=5200 |
|---|---|---|---|---|
| Unnorm. | 1 | 89.1 ± 1.9 | 82.2 ± 0.2 | 77.5 ± 0.0 |
| 3 | 95.4 ± 0.8 | 59.8 ± 0.7 | 68.2 ± 0.4 | |
| 1/μ_\max | 1 | 88.8 ± 1.9 | 98.9 ± 0.1 | 65.4 ± 0.2 |
| 3 | 95.8 ± 0.7 | 99.9 ± 0.0 | 81.5 ± 1.3 | |
| R2L | 1 | 87.0 ± 1.2 | 96.6 ± 0.1 | 96.5 ± 0.0 |
| 3 | 88.4 ± 1.0 | 95.4 ± 0.2 | 95.9 ± 0.0 | |
| R2L(rw) | 1 | 94.6 ± 0.6 | 99.9 ± 0.0 | 99.7 ± 0.0 |
| 3 | 95.9 ± 0.7 | 99.9 ± 0.0 | 99.8 ± 0.0 |
Table B: False positive rate (FPR) and false negative rate (FNR) for the filters. Lower is better.
| Operator | K | n=240 FPR | n=240 FNR | Irreg. n=5200 FPR | Irreg. n=5200 FNR |
|---|---|---|---|---|---|
| Unnorm. | 1 | 35.7 ± 0.4 | 0.0 ± 0.0 | 44.9 ± 0.1 | 0.2 ± 0.0 |
| 3 | 79.5 ± 1.3 | 1.0 ± 0.0 | 63.1 ± 0.7 | 0.6 ± 0.0 | |
| 1/μ_\max | 1 | 1.1 ± 0.1 | 1.2 ± 0.3 | 0.3 ± 0.1 | 68.8 ± 0.3 |
| 3 | 0.3 ± 0.0 | 0.0 ± 0.0 | 0.0 ± 0.0 | 37.0 ± 2.7 | |
| R2L | 1 | 2.3 ± 0.1 | 4.6 ± 0.2 | 2.6 ± 0.0 | 4.4 ± 0.0 |
| 3 | 8.5 ± 0.3 | 0.8 ± 0.0 | 7.2 ± 0.0 | 1.1 ± 0.0 | |
| R2L(rw) | 1 | 0.1 ± 0.0 | 0.1 ± 0.0 | 0.2 ± 0.0 | 0.4 ± 0.0 |
| 3 | 0.2 ± 0.0 | 0.0 ± 0.0 | 0.3 ± 0.0 | 0.0 ± 0.0 |
We see that R2L normalization outperforms the unnormalized filters for transferability in this task. This happens as an unnormalized filter extrapolates the frequency response to unseen eigenvalues, where it disproportionately scales up the outputs compared to the learned threshold. R2L also outperforms the 1/\mu_\max normalization when the underlying simplicial complex is irregular. This is because the irregularity is caused by a dense, local cluster, and 1/\mu_\max disproportionately shrinks the filter outputs, causing the outputs to remain below the threshold. This is visible in Table B. The performance of R2L(rw) comes from the fact that it changes the basis of the signal to be compatible with the edge-signals, which is discussed in more detail in Appendix D.2.
Normalization guarantees that the spectra of di"erent complexes lie in the same interval, but this isn't enough to claim that learned filters transfer: the informative eigenspaces can still be at di"erent frequencies. The book-complex example in Figure 2 is a hand-picked illustration, not evidence.
We agree with the reviewer's comment that the informative eigenspaces might be at the different components, and it is a strong assumption to make that the informative frequencies remain identical for every frequency. Our argument is not that normalized operators always transfer, but that a shared spectral scale is important if transferability is desired.
When a filter is learned on a fixed domain, we only care about its action on the eigenvalues of the domain. When it is transferred to another domain with different eigenvalues, the polynomial will be acting extrapolating to unseen frequencies, which can have negative impact on the performance, and this view is now demonstrated in Figure 5, Section 5.3. Especially since a polynomial filter will be strictly increasing or decreasing, when the spectral scales differ too much, this can have drastic impact on the scale of the output. Normalization mitigates this, also shown by our new experiments. Book complexes are a theoretically tractable example to demonstrate our point, and hopefully the experiments now complement the picture.
We thank the reviewer for pointing this out, and we have changed our wording in the revision, so that our claim is clearer.
Official Comment by Authors - Part 3:
Linear Models, Skew-Symmetry
The theory doesn't directly apply to the proposed model. The analysis considers linear dynamics with only the first-order term and constrained (symmetric or skew-symmetric) weights, while R2L-DCN uses nonlinearities, unconstrained weights, and second-order Chebyshev terms
We like to emphasize that, in this work, we proposed DCNs as a general learning framework, providing both theoretical and empirical contributions. Similar to many previous works including di Giovanni et al., Deep Linear Networks [R2], our theoretical framework considers linear networks and leaves the analysis of networks with non-linear activations for a future study. This is the setting in the Prop. 4.1/4.2. Our empirical analysis on the other hand (e.g. Tab. 1, 2 & 3), deploys non-linear DCNs with normalized operators to practical problems. We kindly ask the reviewer to consider our contributions (DCN framework, linear theoretical analysis and empirical non-linear DCNs) as a whole when making their assessment.
What is specific to Dirac?
Proposition 4.1 holds for any symmetric propagation operator with skew-symmetric weights, so it's unclear what is specific to the Dirac formulation // What is specific to the Dirac formulation in Proposition 4.1?
We agree with the reviewer that any symmetric propagator with skew-symmetric weight matrix gives energy preservation in GNNs, which is explored in Gravina et al. We do not claim the use of skew-symmetry is novel and unique, but the origins of the skew-symmetry have different implications in coupled dynamics.
For a skew-symmetric weight matrix, the evolution of the linear term can be equivalently written as:
which now acts on features with a symmetric weight matrix. Such construction differs from the previous cases in the following ways:
- Skew-symmetry comes from the structure. It is the bi-partite structure of vertices and edges that create the skew-symmetry and not the skew-symmetry of the weights exactly. Now the structure is described by a skew-symmetric analogue of the Dirac operator (call ). Furthermore, the linear dynamics generated by , that is , can be differentiated again to get , that is the rank-wise wave equation. Therefore intuitively, this can be viewed as a first-order way to describe the wave equation. This is the key difference to having skew-symmetric weights. The relationship between how this affects the SNN dynamics is left for future work.
- Dirac operator enforces different structure on . For rank-wise interactions with a 1-channel signals, the only way we can act with a skew-symmetric matrix is if , which cannot learn anything. Similarly for a -channel signal, we enforce the diagonals to be , meaning a feature update cannot be informed by the feature itself. With the Dirac operator, need not satisfy any structure, it can be any matrix. The structure we enforce is that the vertex edge map is the transpose (times ) of the edge vertex. This is a different structure enforced, special to the Dirac case.
- Graded Signal. The vertex and edge signal spaces are assumed to be different kinds, as they will hold different types of information (scalar, flow, circulation...). This differs the Dirac case from a bi-partite graph. In a bi-partite graph, often the signal spaces of both parts are the same, therefore the signal matrix does not necessarily have a block-diagonal structure.
These three points make the skew-symmetry with the Dirac operator fundamentally different than skew-symmetric weight matrices in any rank-wise propagation, and we believe that these create many interesting research directions on the properties of coupled dynamics. We have added this discussion in Appendix E.
Official Comment by Authors - Part 1:
We sincerely thank the reviewer for their detailed feedback and constructive suggestions. We are especially encouraged that the reviewer views our contributions as clean and natural, and that the significance of retaining performance with depth is acknowledged by the reviewer. We address the concerns below:
Novelty
We appreciate the reviewer for bringing [R1] to our attention, and we have cited it in the revised work. We point out the main differences below:
We believe developing new spectral learning methods on simplicial complexes is a crucial contribution, and our formulation can guide next steps such as non-linear filters which respect the coupling on simplicial complexes. Our work also makes it explicit that different kinds of signals may lie on different ranks, and what nuances a graded domain brings when we develop spectral methods.
We agree that the R2L-DCN architecture can be viewed as a replacement of the unnormalized operators with their R2L-normalized counterparts in SCCNN, but we also believe that this is a reductive view of our contributions. The main contribution of the paper is not the R2L-DCN architecture, it is only an instantiation of our framework, which explores:
With this perspective, [R1] and SCCNNs become a special case of our broad framework. We hope one can appreciate the significance of such a framework to compare SNNs in a systematic manner, to see what choices are useful in which tasks. In our new filter transferability experiments, we also use the random-walk Dirac operators which fit into our framework with a simple pre-/post-processing step, showing the flexibility of our framing.
Furthermore, a significant part of our contribution is the theoretical analysis of the asymptotic behaviour of the Dirichlet energy and the Rayleigh quotient with the linear Dirac evolution, and characterizing its possible limits with symmetric and skew-symmetric weights. To the best of our knowledge, we provide first such analysis where the Dirac operator interacts with the weight matrices acting on the features. Our result shows that with (skew-)symmetric weight matrices, regardless of the learned weights, the Dirac operator tends to act a a high-pass filter. This result is novel compared to previous analysis on coupled SNNs and analogous results in GNNs, and exploring coupled dynamics is crucial for extending notions of oversmoothing to simplicial complexes.
Finally, we also discuss the interpretation of normalization from a spectral-scale perspective, which has not been made clear in the SNN case. We believe that such discussions will be helpful to decide which architecture should be used for certain tasks.
We hope to have conveyed that our work is indeed novel, goes beyond [R1] and SCCNNs in our spectral framing, theoretical analysis, and practical implications. We believe that our exposition, now strengthened by this discussion is crucial to understand the use cases of topological deep learning by emphasizing the significance of the choices we make while designing the architectures.
We want to clarify that, our argument in the paper is not that inter-rank coupling is always beneficial. We introduce a framework which encompasses the possible choices that can be made for SNN architectures, informing a broad, task-dependent design of simplicial complex networks.